Nvidia has deepened its partnership with ElevenLabs, the AI voice company, at the New York summit that Jensen Huang was supposed to headline. The chipmaker said on 11 November that its Nemotron models are now available inside ElevenAgents, ElevenLabs’ platform for building voice agents, in a post at 21:17 UTC, or 16:17 in New York. Kari Briski, one of Nvidia’s generative AI software leaders, represented the company on stage instead of its chief executive. Less than an hour later, The New York Times reported a second relationship involving Huang, and by the evening the Wall Street Journal said Nvidia’s board was preparing to remove him.
Huang had been booked for a fireside chat with ElevenLabs chief executive Mati Staniszewski. Business Insider reported on 9 November that his name had disappeared from the speaker list, a day after he admitted to an affair in a letter to staff. The agenda was then rewritten to give Staniszewski the closing keynote instead. Huang also stepped back from a separate gathering of chief executives this week, according to reports. The partnership went ahead without him, which says a lot about how Nvidia is trying to keep its business moving while its leadership is in question.
A cheaper Nvidia brain for ElevenLabs’ voice agents
The centrepiece of the deal is Nemotron 3.5 Super VL, one of Nvidia’s open models, which ElevenLabs customers can now choose inside ElevenAgents. According to the ElevenLabs announcement, the model reached leading task-completion accuracy on the company’s customer-support benchmark at 70% lower cost than other leading open models. Its reasoning can also be switched off for simple tasks, which ElevenLabs says makes it one of the three fastest models on the platform. For businesses running thousands of calls a day, that combination of price and speed is the real selling point.
The bigger shift in this deal is about control over data. Companies can post-train Nemotron models on their own data and run the whole system on their own servers, rather than sending conversations to an outside cloud. That matters for banks, hospitals and governments, which often cannot let sensitive calls leave their own buildings. ElevenLabs is also adding 15 megawatts of Nvidia Vera Rubin computing capacity to its existing Blackwell Ultra systems, a sign of how much demand it expects.
The first public example comes from Poland, the home country of ElevenLabs’ co-founders. A voice agent running on-premises with Nvidia Confidential Computing will call patients who only have landlines, reminding them of appointments booked through the national e-Health system. That system handles about 40 million appointments a year, and ElevenLabs says missed appointments run between 10% and 20%. The two companies are also exploring voice AI that runs directly on robots, industrial machines and vehicles, although that work is still at an early stage.
Huang’s words made the press release, but he never made the stage
The ElevenLabs announcement still carries a quote from Huang, who said the aim is to “make talking the fastest way to get things done”. Staniszewski said businesses need to operate securely as voice agents “become critical infrastructure”. It is an awkward split screen for Nvidia, with its founder’s vision in the release and his seat empty at the event. The relationship itself runs deep, and it did not depend on one appearance.
The two companies have worked together since 2022, when ElevenLabs trained its first text-to-speech model on Nvidia GPUs. Since 2024, Nvidia has used ElevenLabs to produce voiceovers for more than 50 videos from its biggest events, including translated versions of Huang’s own keynotes. Nvidia then became a strategic investor in ElevenLabs in 2025. Few partners are as closely tied to the Huang era at Nvidia, which made his absence from the summit all the more noticeable.
The timing of the summit could hardly have been worse for Nvidia. Its independent directors had already brought in former Manhattan prosecutor Joon Kim to lead their investigation, and investors had wiped about $452 billion off Nvidia’s value in the first two sessions after Huang’s letter. The ElevenLabs news was the kind of steady partnership update that would normally pass quietly. This week, it landed in the middle of a boardroom crisis.
From Darth Vader to talking NPCs, ElevenLabs is already in your games
Voice AI is not just for call centres, and ElevenLabs has been part of its move into games. When Epic Games brought an AI-voiced Darth Vader to Fortnite in 2025, it said the voice used ElevenLabs’ Flash v2.5 model alongside Google’s Gemini, with permission from James Earl Jones’ estate. The actors’ union SAG-AFTRA objected to that use of AI to replace performers, at a time when its video game members were on strike over AI protections. That strike ended in 2025 with new rules on consent and pay for digital replicas of performers’ voices. The debate has not gone away, even as Epic has since given talking AI voices to 36 Fortnite characters for creators to use on their islands.
ElevenLabs has been courting game developers directly too, publishing guides on giving Unity characters a voice through ElevenAgents. Nvidia, meanwhile, has pushed its own ACE technology for conversational game characters, which uses small language models, including a Nemotron model shown in the mech game Mecha BREAK. With Nemotron now inside ElevenAgents, a studio could build a talking character from both companies’ tools on its own hardware. That could appeal to developers who want to keep unreleased scripts and story details off outside servers.
For players, the result could be characters that answer questions, react to what happens in a match and speak in many languages. For voice actors, the same tools raise hard questions about consent and pay, which is why unions are watching every deal like this one. ElevenLabs used its New York summit to launch new memory and routine features for its agents, which could just as easily power a game’s support desk as a bank’s. The company says its partner programme now covers more than 550 partners in 44 countries.
